def AddCnnLayers(config_lines, cnn_layer, cnn_bottleneck_dim, cepstral_lifter, config_dir, feat_dim, splice_indexes=[0], ivector_dim=0):
cnn_args = ParseCnnString(cnn_layer)
num_cnn_layers = len(cnn_args)
# We use an Idct layer here to convert MFCC to FBANK features
common_lib.write_idct_matrix(feat_dim, cepstral_lifter, config_dir.strip() + "/idct.mat")
prev_layer_output = {'descriptor': "input",
'dimension': feat_dim}
prev_layer_output = nodes.AddFixedAffineLayer(config_lines, "Idct", prev_layer_output, config_dir.strip() + '/idct.mat')
list = [('Offset({0}, {1})'.format(prev_layer_output['descriptor'],n) if n != 0 else prev_layer_output['descriptor']) for n in splice_indexes]
splice_descriptor = "Append({0})".format(", ".join(list))
cnn_input_dim = len(splice_indexes) * feat_dim
prev_layer_output = {'descriptor': splice_descriptor,
'dimension': cnn_input_dim,
'3d-dim': [len(splice_indexes), feat_dim, 1],
'vectorization': 'yzx'}
for cl in range(0, num_cnn_layers):
prev_layer_output = AddConvMaxpLayer(config_lines, "L{0}".format(cl), prev_layer_output, cnn_args[cl])
if cnn_bottleneck_dim > 0:
prev_layer_output = nodes.AddAffineLayer(config_lines, "cnn-bottleneck", prev_layer_output, cnn_bottleneck_dim, "")
if ivector_dim > 0:
iv_layer_output = {'descriptor': 'ReplaceIndex(ivector, t, 0)',
'dimension': ivector_dim}
iv_layer_output = nodes.AddAffineLayer(config_lines, "ivector", iv_layer_output, ivector_dim, "")
prev_layer_output['descriptor'] = 'Append({0}, {1})'.format(prev_layer_output['descriptor'], iv_layer_output['descriptor'])
prev_layer_output['dimension'] = prev_layer_output['dimension'] + iv_layer_output['dimension']
return prev_layer_output
def AddCnnLayers(config_lines, cnn_layer, cnn_bottleneck_dim, cepstral_lifter, config_dir, feat_dim, splice_indexes=[0], ivector_dim=0):
cnn_args = ParseCnnString(cnn_layer)
num_cnn_layers = len(cnn_args)
# We use an Idct layer here to convert MFCC to FBANK features
common_lib.write_idct_matrix(feat_dim, cepstral_lifter, config_dir.strip() + "/idct.mat")
prev_layer_output = {'descriptor': "input",
'dimension': feat_dim}
prev_layer_output = nodes.AddFixedAffineLayer(config_lines, "Idct", prev_layer_output, config_dir.strip() + '/idct.mat')
list = [('Offset({0}, {1})'.format(prev_layer_output['descriptor'],n) if n != 0 else prev_layer_output['descriptor']) for n in splice_indexes]
splice_descriptor = "Append({0})".format(", ".join(list))
cnn_input_dim = len(splice_indexes) * feat_dim
prev_layer_output = {'descriptor': splice_descriptor,
'dimension': cnn_input_dim,
'3d-dim': [len(splice_indexes), feat_dim, 1],
'vectorization': 'yzx'}
for cl in range(0, num_cnn_layers):
prev_layer_output = AddConvMaxpLayer(config_lines, "L{0}".format(cl), prev_layer_output, cnn_args[cl])
if cnn_bottleneck_dim > 0:
prev_layer_output = nodes.AddAffineLayer(config_lines, "cnn-bottleneck", prev_layer_output, cnn_bottleneck_dim, "")
if ivector_dim > 0:
iv_layer_output = {'descriptor': 'ReplaceIndex(ivector, t, 0)',
'dimension': ivector_dim}
iv_layer_output = nodes.AddAffineLayer(config_lines, "ivector", iv_layer_output, ivector_dim, "")
prev_layer_output['descriptor'] = 'Append({0}, {1})'.format(prev_layer_output['descriptor'], iv_layer_output['descriptor'])
prev_layer_output['dimension'] = prev_layer_output['dimension'] + iv_layer_output['dimension']
return prev_layer_output